Qwen3.5-27B-Omnimerge-GGUF

GGUF quantizations of ManniX-ITA/Qwen3.5-27B-Omnimerge — a 3-way Task Arithmetic weight-space merge of three Qwen3.5-27B reasoning-distilled fine-tunes.

This merge outperforms its best source model (Claude-4.6-Opus-Reasoning-Distilled) across all tested benchmarks: +8 pp on GPQA Diamond reasoning, +3.7 pp on HumanEval, and comparable MBPP.

All quants made with imatrix using calibration data v5.

Benchmark Results (Q6_K)

Benchmark	Omnimerge	Claude-distill (best source)	Delta
GPQA Diamond (198q, flex)	61.11%	53.03%	+8.08 pp
HumanEval pass@1	79.88%	76.22%	+3.66 pp
MBPP pass@1	71.80%	71.20%	+0.60 pp

Available Quantizations

Quantization	File	Size
Q8_0	merged_omnimerge-Q8_0.gguf	26.63 GB
Q6_K_L	merged_omnimerge-Q6_K_L.gguf	21.14 GB
Q6_K	merged_omnimerge-Q6_K.gguf	20.57 GB
Q5_K_L	merged_omnimerge-Q5_K_L.gguf	18.64 GB
Q5_K_M	merged_omnimerge-Q5_K_M.gguf	17.91 GB
Q5_K_S	merged_omnimerge-Q5_K_S.gguf	17.40 GB
Q4_K_L	merged_omnimerge-Q4_K_L.gguf	16.29 GB
Q4_1	merged_omnimerge-Q4_1.gguf	15.91 GB
Q4_K_M	merged_omnimerge-Q4_K_M.gguf	15.41 GB
IQ4_NL	merged_omnimerge-IQ4_NL.gguf	14.72 GB
Q4_K_S	merged_omnimerge-Q4_K_S.gguf	14.52 GB
Q4_0	merged_omnimerge-Q4_0.gguf	14.41 GB
IQ4_XS	merged_omnimerge-IQ4_XS.gguf	14.05 GB
Q3_K_XL	merged_omnimerge-Q3_K_XL.gguf	13.42 GB
Q3_K_L	merged_omnimerge-Q3_K_L.gguf	13.36 GB
Q3_K_M	merged_omnimerge-Q3_K_M.gguf	12.39 GB
IQ3_M	merged_omnimerge-IQ3_M.gguf	11.72 GB
Q3_K_S	merged_omnimerge-Q3_K_S.gguf	11.24 GB
IQ3_XS	merged_omnimerge-IQ3_XS.gguf	11.15 GB
IQ3_XXS	merged_omnimerge-IQ3_XXS.gguf	10.42 GB

Skipped Quantizations (failed sanity check)

The following 2-bit quantizations were attempted but failed the sanity check (3 capital city questions answered incorrectly or incoherently). At 2-bit precision on a 27B model, too much information is lost for reliable output. These quants are intentionally not published:

• Q2_K_L, Q2_K, IQ2_M, IQ2_S, IQ2_XS, IQ2_XXS

Recommended Usage

llama-server -m Qwen3.5-27B-Omnimerge-Q6_K.gguf -c 32768 -ngl 99 \
    --jinja --reasoning-format deepseek --reasoning-budget 16384 \
    --temp 0.6 --top-p 0.95 --top-k 20 --dry-multiplier 0.5

For code tasks, use without --jinja --reasoning-format deepseek (plain completions mode).

Source Models

Source	Weight	Focus
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled	0.40	Claude 4.6 Opus reasoning distillation
ValiantLabs/Qwen3.5-27B-Esper3.1	0.35	Code / DevOps specialist
DavidAU/Qwen3.5-27B-Gemini3-Pro-High-Reasoning-Compact-Thinking	0.25	Gemini 3 Pro reasoning, compact thinking

Base: Qwen/Qwen3.5-27B

See the model card for full methodology, evaluation details, and the custom merger script.

License

Apache-2.0